ready for generic prompt

.env.example

@@ -2,6 +2,7 @@ MODEL_NAME=Qwen/Qwen2-7B-Instruct
 
 BATCH_SIZE=2
 MAX_NEW_TOKENS=300
+USING_CHAT_TEMPLATE=true
 
 HF_TOKEN=
 

.gitignore

@@ -151,3 +151,4 @@ dmypy.json
 /llama.cpp
 /llama-factory/config/models
 /codedrive
+temp.csv

llm_toolkit/eval_rpp.py

@@ -27,6 +27,7 @@ data_path = os.getenv("DATA_PATH")
 results_path = os.getenv("RESULTS_PATH")
 batch_size = int(os.getenv("BATCH_SIZE", 1))
 use_english_datasets = os.getenv("USE_ENGLISH_DATASETS") == "true"
+using_chat_template = os.getenv("USING_CHAT_TEMPLATE") == "true"
 max_new_tokens = int(os.getenv("MAX_NEW_TOKENS", 2048))
 start_repetition_penalty = float(os.getenv("START_REPETITION_PENALTY", 1.0))
 end_repetition_penalty = float(os.getenv("END_REPETITION_PENALTY", 1.3))
@@ -63,7 +64,7 @@ if is_cuda:
     print(f"(2) GPU = {gpu_stats.name}. Max memory = {max_memory} GB.")
     print(f"{start_gpu_memory} GB of memory reserved.")
 
-datasets = load_translation_dataset(data_path, tokenizer)
+datasets = load_translation_dataset(data_path, tokenizer, using_chat_template=using_chat_template)
 
 if len(sys.argv) > 1:
     num = int(sys.argv[1])
@@ -82,7 +83,7 @@ def on_repetition_penalty_step_completed(model_name, predictions):
         predictions,
     )
 
-    metrics = calc_metrics(datasets["test"]["english"], predictions, debug=True)
+    metrics = calc_metrics(datasets["test"]["english"], predictions, datasets["test"]["chinese"], debug=True)
     print(f"{model_name} metrics: {metrics}")
 
 

llm_toolkit/translation_utils.py

@@ -118,7 +118,7 @@ def get_few_shot_prompt(dataset, num_shots=5):
     return translation_prompt
 
 
-def load_translation_dataset(data_path, tokenizer=None, num_shots=0, for_openai=False):
+def load_translation_dataset(data_path, tokenizer=None, num_shots=0, for_openai=False, using_chat_template=True):
     train_data_file = data_path.replace(".tsv", "-train.tsv")
     test_data_file = data_path.replace(".tsv", "-test.tsv")
 
@@ -187,7 +187,8 @@ def load_translation_dataset(data_path, tokenizer=None, num_shots=0, for_openai=
                 else:
                     prompt = tokenizer.apply_chat_template(
                         messages, tokenize=False, add_generation_prompt=True
-                    )
+                    ) if using_chat_template else prompt
+                    
                     prompts.append(prompt)
                     texts.append(prompt + output + tokenizer.eos_token)
 

requirements.txt

@@ -13,7 +13,7 @@ packaging
 langchain_openai==0.1.13
 wandb==0.17.6
 transformers==4.43.3
-bitsandbytes==0.43.3
+bitsandbytes #==0.43.3
 sentencepiece==0.1.98
 einops==0.8.0
 accelerate==0.32.0

scripts/eval-4gpu.sh

@@ -31,7 +31,12 @@ export LOAD_IN_4BIT=true
 export MAX_NEW_TOKENS=2048
 export START_REPETITION_PENALTY=1.0
 export END_REPETITION_PENALTY=1.1
-export RESULTS_PATH=results/mac-results_rpp_with_mnt_2048.csv
+
+export USING_CHAT_TEMPLATE=false
+export RESULTS_PATH=results/mac-results_rpp_with_mnt_2048_generic_prompt.csv
+
+# export USING_CHAT_TEMPLATE=true
+# export RESULTS_PATH=results/mac-results_rpp_with_mnt_2048.csv
 
 ./scripts/eval-rpp.sh shenzhi-wang Llama3.1-70B-Chinese-Chat checkpoint-210
 

scripts/eval-mac.sh

@@ -43,14 +43,19 @@ export RESULTS_PATH=results/mac-results_fine_tuned.csv
 export MAX_NEW_TOKENS=2048
 export START_REPETITION_PENALTY=1.0
 export END_REPETITION_PENALTY=1.1
-export RESULTS_PATH=results/mac-results_rpp_with_mnt_2048.csv
 
-# ./scripts/eval-rpp.sh internlm internlm2_5-7b-chat checkpoint-140
+export USING_CHAT_TEMPLATE=false
+export RESULTS_PATH=results/mac-results_rpp_with_mnt_2048_generic_prompt.csv
 
-# ./scripts/eval-rpp.sh Qwen Qwen2-7B-Instruct checkpoint-105
+# export USING_CHAT_TEMPLATE=true
+# export RESULTS_PATH=results/mac-results_rpp_with_mnt_2048.csv
 
-# ./scripts/eval-rpp.sh shenzhi-wang Mistral-7B-v0.3-Chinese-Chat checkpoint-70
+./scripts/eval-rpp.sh internlm internlm2_5-7b-chat checkpoint-140
 
-# ./scripts/eval-rpp.sh shenzhi-wang Llama3.1-8B-Chinese-Chat checkpoint-105
+./scripts/eval-rpp.sh Qwen Qwen2-7B-Instruct checkpoint-105
+
+./scripts/eval-rpp.sh shenzhi-wang Mistral-7B-v0.3-Chinese-Chat checkpoint-70
+
+./scripts/eval-rpp.sh shenzhi-wang Llama3.1-8B-Chinese-Chat checkpoint-105
 
 ./scripts/eval-rpp.sh microsoft Phi-3.5-mini-instruct checkpoint-210